Nvidia puts a 397 MB humanoid walking policy in the browser

Nvidia has published Metatron v1 browser exports on Hugging Face: two ONNX files, each close to 400 MB, that run a walking and recovery policy for the 29-joint Unitree G1 humanoid inside a web page.

Two files that are not interchangeable

The repository holds demo/model.onnx (397,088,354 bytes) and racer/network.onnx (396,370,481 bytes). According to the model card, both carry the same learned weights but expose different interfaces, and the card says they are “not interchangeable.”

The demo file is a single-robot pipeline that includes observation processing and action conversion. As the card describes it, the network receives the robot’s present state together with a short trail of eight earlier state readings, the previous raw actions, the terrain height under the robot, and partly masked upcoming targets for pose, style and velocity. A recovery_enabled switch is also an explicit input. It returns joint PD targets, stiffness and damping values, plus a raw_action output that is fed back as history.

The Racer file is a six-robot batched network for the G1 Racer game. It computes observations outside the neural graph, so it needs that application’s own preprocessing. Nvidia says the Racer export changes graph organization and batching, not the weights, and warns against changing its batch size without rebuilding and validating it.

Nvidia describes the model as a conditional generative control policy with a gated recovery adapter, run deterministically in FP32. In the browser, MuJoCo compiled to WebAssembly handles physics and ONNX Runtime Web handles inference. WebGPU is recommended; the CPU/WebAssembly fallback does not guarantee real-time operation with six robots.

The only accuracy figure is a numerical one: recorded native-to-ONNX checks on the demo export show maximum absolute errors below 3.1e-6, with recovery on and off. Nvidia stresses these are export checks, not task-success or real-robot benchmarks.

Who can use it, and what is missing

The files can be downloaded anonymously, and the model is covered by the NVIDIA Open Model License Agreement. The files alone are not a working demo. Physics, observations, task goals and control logic come from the matching browser applications, and the card links no public address for either the Metatron demo or G1 Racer. Motion clips, robot meshes, game assets and the separate SAMP scene-interaction policy are not included.

Nor can anyone retrain it from this repository. There is no training dataset and no resumable checkpoint. The card points to ProtoMotions, Nvidia’s research framework, as the reference. Developers who want to deploy elsewhere, such as other simulators or a real G1, would have to match the paired application’s joint ordering, quaternion convention, history ordering, normalization and control rate exactly.

Nvidia also advises pinning downloads to a repository commit rather than main, and never embedding a Hugging Face token in client-side code, since a public browser demo needs anonymous file access.

A safety line drawn around a robot that walks

The caution in this card is stronger than in most model releases, and it matters because the underlying approach is not confined to a screen. Nvidia’s ProtoMotions repository says a general tracking policy trained there has transferred zero-shot to a real G1, deployed through the RoboJuDo framework. The Metatron card keeps that history at arm’s length: compatibility with real robots is “not a safety certification,” and Nvidia does not claim these exact files were validated on hardware.

The card adds that the model is not a collision-avoidance system or a safety controller. It names independent validation, joint and torque limits, supervised testing and an emergency stop as requirements for hardware use. Recovery, balance and tracking can fail outside the evaluated states, terrain and commands, according to the card. It also separates game mechanics, such as launch assists, low-gravity zones and respawns, from anything the policy or a physical robot can do.

The licensing is layered too. ProtoMotions code is Apache-2.0, while Nvidia’s pretrained models there fall under the NVIDIA Open Model License unless a card says otherwise, according to the repository. The Metatron card adds that the model license grants no rights to motion datasets, recovery clips, meshes or game assets, and that permission to show a clip in a demo is not permission to redistribute it as a dataset.

The pattern fits Nvidia’s other recent humanoid release. Its MotionBricks card describes a motion generator rather than a robot controller, says hardware use needs a separate tracking controller, and omits the demo motion bank from the G1 package. For readers following robotics foundation models, such as Google’s Gemini Robotics ER 2, Metatron is a narrower item: a low-level motor policy, not a reasoning model.

What the source does not establish is how well the policy performs. There are no task-success rates, no comparison with earlier Nvidia controllers and no hardware results, and the browser demos themselves are not linked from the card. Until those appear, Metatron is best read as a downloadable, inspectable piece of a larger Nvidia humanoid stack rather than a ready-to-run robot brain.